Reproduce “Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, Jiajun Liang
A Joint Token Pruning & Squeezing (TPS) module is introduced to aggressively compress Vision Transformers by intelligently preserving information from pruned tokens. This method improves accuracy by 1-6% over existing baselines under high compression ratios, enabling more efficient ViT deployment.
Key insights
- Discarded tokens in pruning methods often retain valuable, complementary information that, if preserved, could improve classification accuracy, especially for challenging cases.
- Existing token pruning and reorganization methods suffer from information loss or "feature collapse," limiting their ability to achieve very aggressive compression ratios without significant performance drops.
- Intelligently transferring features from pruned tokens into relevant reserved tokens can effectively mitigate information loss, allowing for more robust and aggressive Vision Transformer compression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.